Bruce Momjian wrote:
> Bruce Momjian wrote:
> > > It does seem possible that that could happen, but I'm not sure exactly
> > > what would be causing autovacuum to fire in the first place. It
> > > wouldn't have to be triggered by the anti-wraparound machinery - if
> > > the table appeared to be in need of vacuuming, then we'd vacuum it,
> > > discover that is was empty, and update relfrozenxid. Hmm... could it
> > > fire just because the table has no stats? But if that were the case
> > > you'd think we'd be seeing this more often.
> >
> > Well, autovacuum=off, so it should only run in freeze mode, and I can't
> > see how that could happen. I am thinking I have to study autovacuum.c.
> >
> > I wonder if datfrozenxid could be incremented because the database is
> > originally empty. It would just need to scan pg_class, not actually
> > vacuum anything. I wonder if we do that. The bottom line is I am
> > hanging too much on autovacuum_freeze_max_age causing autovacuum to do
> > nothing.
>
> What if we allow autovacuum_max_workers to be set to zero; the current
> minimum is one.
I can think of one case where autovacuum_freeze_max_age would be
insufficient. If you set autovacuum_freeze_max_age in the old cluster
to 2B, and you had a database that was near that limit, the tables
created by pg_upgrade's --schema-only restore might create enough new
transactions to cause autovacuum to run in freeze mode. While I think
it is unlikely that is the cause of the problem report, it is enough for
me to discount using autovacuum_freeze_max_age to disable autovacuum
freeze.
I will work on code to allow autovacuum_max_workers to be set to zero in
HEAD and 9.0, and have pg_upgrade us that. I think the maintenance
overhead of an invisible variable is too much.
-- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB
http://enterprisedb.com
+ It's impossible for everything to be true. +
